Section 1
Deterministic and probabilistic, in plain terms
A rule engine moves a record when a field changes. It has no opinion, no variance, and no capacity to handle an input it was not told about. Ninety percent of what businesses call automation is this, and it is a good technology. A generative model predicts likely continuations. It can read an email nobody standardized, infer intent, summarize a call, or draft a reply in your house style. It can also produce a confident answer that is wrong, and the wrongness looks exactly like the rightness, which is what makes it different from a broken rule. So the design principle is not one or the other. It is a deterministic spine with generative steps at the points where the input is unstructured or the output is a first draft. The rules carry the guarantees. The model carries the ambiguity.
Section 2
The jobs that only a generative step can do
Reading, in the real sense. A supplier invoice in an unfamiliar layout, a claim described in a customer's own words, a contract clause that needs to be found rather than located by field name. Classifying against fuzzy categories. Whether a message is a complaint or a query, whether a lead matches a profile, whether a support ticket concerns billing or product, when the boundaries are judgment rather than keywords. Drafting. Proposals, replies, documentation, release notes. The economics here are the strongest in the whole category, because a first draft is slow to write and fast to correct, and the human keeps final authorship. Summarizing for a decision. A weekly digest of what changed, with the exceptions surfaced. Not a report generator, but a filter that hands a person the twelve things worth their attention out of four hundred.
Section 3
Where to place the probabilistic step
The placement rule is short: a generative step is safe at the entrance of a process, where a human or a validator sits between it and any consequence, and risky in the middle, where downstream systems will treat its output as fact. The model below maps the placements against risk. For where this sits in the wider market direction, see [Top AI Automation Trends for 2026 and Beyond](/blog/top-ai-automation-trends-for-2026-and-beyond).
Section 4
Wiring it into a pipeline that already works
Do not replace the existing rules. Wrap them. The workflow that already routes tickets keeps routing tickets, and the generative step supplies the classification the rules used to receive from a person. Constrain the output shape. A free-text answer is difficult to act on. Ask for the specific fields you need, validate them against your schema, and treat anything failing validation as an exception rather than as a slightly odd success. That single practice removes a large share of production problems. Give it the context it needs rather than assuming it knows. A model that has not been shown your product catalogue, your pricing rules or your last thirty approved replies is inventing a plausible version of your business. And keep the human in the loop where it earns its cost: on the first few weeks of everything, and permanently on anything consequential.
Section 5
Non-determinism and the checks it demands
The uncomfortable property is that the same input can produce a different output tomorrow, because the model changed, the prompt drifted, or the sampling differed. Traditional testing assumes stability, so it needs supplementing rather than replacing. Keep an evaluation set of real examples with known correct answers, including the awkward ones, and run it whenever anything changes. Log inputs, outputs and escalations so a bad result can be reconstructed. Watch for silent degradation, which is the characteristic failure here: nothing errors, the answers just get slightly worse. The governance questions are the same ones any risk framework asks. What can this system read, what can it change, what must it never decide alone, and whose name is on it when a customer is affected. Disclose where AI is involved. A case study of that discipline in practice is [Startup Success: How AI Automation Transformed Our Business](/blog/startup-success-how-ai-automation-transformed-our-business).
Section 6
Measuring a step that is right most of the time
Accuracy on a held-out set is the base metric, but it is not the business one. Track the escalation rate, meaning how often the output goes to a human. Track the override rate, meaning how often that human changed it. Track the cost of a miss, separated into cheap misses and expensive ones. A generative step that is correct eighty percent of the time can be excellent or unusable, and the deciding factor is not the eighty. It is whether the twenty percent is caught cheaply, and whether being wrong in that workflow costs a correction or a customer.